Plan and run agent tasks
Four lessons on the path from an idea to a pushed branch: split the work into issues, check what a change touches, run an agent on one issue, and choose the role prompt it gets.
How the pieces connect
flowchart LR
A["Task Planner"] -->|creates issues| B[("GitHub Issues or Jira")]
B -->|lists tasks| C["Task Implementer"]
C -->|starts| D["claude CLI"]
D -->|pushes| E["Feature branch on origin"]
E -->|Review branch| F["PR Review"]
F -->|Create PR| G["Pull request"]
Task Planner and Task Implementer both read the active task server from ~/.config/robos/settings.json, which you edit in the Task Servers app. The agent in the middle is the Claude Code CLI. Task Implementer starts it and streams its output. It does not create a sandbox, clone the repository or open the pull request. Run an agent on a task lists exactly what it does.
Lessons
| Lesson | Apps used |
|---|---|
| Plan a feature in Task Planner | Task Planner, Task Servers |
| Check impact with the kgraph CLI | Knowledge Graph Explorer (the same graph, as a CLI) |
| Run an agent on a task | Task Implementer, PR Review |
| Choose and customize agent personas | Task Implementer, Task Planner |
Reviewing the branch an agent pushes is covered on the PR Review Theater page.